IJCV paper on learning single-image 3D reconstruction by generative modelling of shape, pose and shading